← All speakers

Diogo Almeida

Conference affiliation: CEO · TypeSafe AI · 2026

Diogo Almeida is co-founder and chief executive of TypeSafe AI, where he is building Jev, an AI model that returns typed decisions and probabilities directly to software. At OpenAI, he coauthored InstructGPT, helping develop the human-feedback training behind conversational AI. His subsequent work tackles a limitation he sees in that approach: a model can produce responses people prefer without being ready to make unattended operational decisions.

On this page

From neural architectures to instruction-following

Diogo’s career spans software engineering at Amazon in 2013–2014 and Stuph.co in 2014, a senior data scientist role at Enlitic in 2014–2017, and research at Google Brain in 2017–2018. He joined OpenAI in 2020 and remained there until 2024.

His early research included Resnet in Resnet, a coauthored dual-stream architecture that generalizes residual networks and conventional convolutional networks. Later, he worked on a different obstacle to adaptable learning systems: optimization methods that perform well on training problems but fail on unfamiliar tasks.

In the 2021 paper A Generalizable Approach to Learning Optimizers, Diogo, Clemens Winter, Jie Tang, and Wojciech Zaremba described a system that learns to adjust optimizer hyperparameters rather than directly update model parameters. The team tested it across modalities absent from training and at substantially larger computational scales. They reported a twofold speedup on ImageNet and a 2.5-fold speedup on a language-modeling task using over five orders of magnitude more compute than the training tasks.

At OpenAI, Diogo coauthored InstructGPT, published in 2022. The team trained GPT-3 on demonstrations of desired responses, collected human rankings of alternative answers, and used reinforcement learning from human feedback, or RLHF, to improve instruction-following. Diogo and Ryan Lowe also developed candidate instruction prefixes for the prompted GPT-3 baseline; Diogo’s achieved the highest validation reward-model score. This supplied a stronger comparison: the study tested human-feedback training alongside an attempt to improve GPT-3’s behavior through prompting alone. On the study’s evaluation prompt distribution, human evaluators preferred responses from a 1.3-billion-parameter InstructGPT model to those from the 175-billion-parameter GPT-3. The result demonstrated how training for user intent could improve preferred behavior without increasing model size; the resulting models still made simple mistakes.

Diogo also helped develop the research behind ChatGPT and coauthored the GPT-4 technical report. His later criticism of preference training grew from participation in the collaborative work that established its usefulness.

Why assistance did not settle automation

In his Jev launch essay, Diogo describes initially thinking chat models might lead toward artificial general intelligence. He became dissatisfied with how little ordinary work their intelligence had automated. He left OpenAI in 2024 and began TypeSafe that August, pursuing models designed for decisions inside software.

His World’s Fair 2026 talk, What’s next after RLHF?, explains the distinction behind that move. Diogo values ChatGPT and coding assistants, while questioning whether their training objectives prepare them for work without continuous human supervision.

  • Assistance versus automation: Answering a customer’s questions and authorizing an expensive action expose a business to different consequences. Diogo distinguishes satisfying someone who can inspect and redirect a response from making a decision in a process running in the background. He wants AI to expand what software can do, alongside making software easier to write.
  • Uncertainty as useful information: Diogo argues that a preference reward model can more easily recognize and penalize hesitation than detect a confident mistake. In his account, this encourages convincing answers and suppresses plausible alternatives. He considers pretrained intelligence valuable; his objection concerns how post-training directs it.
  • Different optimization goals: He distinguishes human preference, verifiable correctness, and calibrated decision-making. For his automation agenda, the model must communicate uncertainty so software can decide whether to act, defer, or gather more information. An answer’s consequences matter alongside its correctness.

His concern also extends to composing model judgments into larger systems. Responding to reports of tool-invocation failures, Diogo suggested excessive on-policy reinforcement learning could help explain difficulties transferring decomposition, abstraction, and state management into language-model workflows. This was a proposed explanation, not an experimental diagnosis of those failures.

Jev: giving up text to make decisions

After two years in stealth, Diogo introduced Jev in September 2026. TypeSafe calls its model family System One Models and its training method Reinforcement Learning for Calibrated Decisions (RLCD). His launch essay describes an automation-focused architecture and a parallel sampler that produces structured outputs together instead of generating text token by token. Giving up arbitrary string generation sacrifices flexibility while constraining the form of the answers software receives.

Jev’s API accepts state and typed questions. Choice selects among supplied options, Score evaluates an ordered rubric, and Noul estimates whether a statement is true. Questions are evaluated independently against the same state, and developers combine the results in code.

The documentation’s startup-pitch example makes that division concrete: ask separately about market size, technical feasibility, and differentiation, then combine the scores with explicit weights. Changing priorities can mean changing a coefficient rather than rewriting a prompt that asks the model to weigh everything at once. The model supplies individual judgments; the program determines their effect on the outcome.

Choice and Score return probability distributions and a confidence statistic derived from them. The confidence documentation illustrates the distinction with three choices: probabilities of 0.6, 0.3, and 0.1 yield confidence of 0.4. That statistic measures how clearly the leading option exceeds an even split; it is not a 40 percent probability that the answer is correct. Developers can use the full distribution or another measure, and should test action thresholds on their own data according to the consequences of mistakes. Type validity constrains an answer’s shape; it does not establish the correctness of the judgment.

Diogo also describes a concrete practice for the team doing this research. In public posts, he explains TypeSafe’s daily lunchtime emotional check-ins: colleagues briefly share how they are feeling while everyone else listens. He rejects the assumption that working hard and caring for teammates are incompatible.